Update app.py

app.py

@@ -6,11 +6,11 @@ from pynvml import *
 nvmlInit()
 gpu_h = nvmlDeviceGetHandleByIndex(0)
 
-title = "RWKV-4 14B fp16 (DEMO, limited to ctxlen 824)"
-desc = '''Links:
-<a href='https://github.com/BlinkDL/ChatRWKV' target="_blank" style="margin:0 1em">ChatRWKV</a>
-<a href='https://github.com/BlinkDL/RWKV-LM' target="_blank" style="margin:0 1em">RWKV-LM</a>
-<a href="https://pypi.org/project/rwkv/" target="_blank" style="margin:0 1em">RWKV pip package</a>
+title = "RWKV-4 14B fp16"
+desc = '''DEMO limited to ctxlen 824, and slow because A10g does not have enough VRAM (some layers are computed on CPU instead). Links:
+<a href='https://github.com/BlinkDL/ChatRWKV' target="_blank" style="margin:0 0.5em">ChatRWKV</a>
+<a href='https://github.com/BlinkDL/RWKV-LM' target="_blank" style="margin:0 0.5em">RWKV-LM</a>
+<a href="https://pypi.org/project/rwkv/" target="_blank" style="margin:0 0.5em">RWKV pip package</a>
 '''
 
 os.environ["RWKV_JIT_ON"] = '1'
@@ -18,7 +18,7 @@ os.environ["RWKV_CUDA_ON"] = '1' # if '1' then use CUDA kernel for seq mode (muc
 
 from rwkv.model import RWKV
 model_path = hf_hub_download(repo_id="BlinkDL/rwkv-4-pile-14b", filename="RWKV-4-Pile-14B-20230213-8019.pth")
-model = RWKV(model=model_path, strategy='cuda fp16 *34 -> cpu fp32')
+model = RWKV(model=model_path, strategy='cuda fp16 *30 -> cpu fp32')
 # model_path = hf_hub_download(repo_id="BlinkDL/rwkv-4-pile-169m", filename="RWKV-4-Pile-169M-20220807-8023.pth")
 # model = RWKV(model=model_path, strategy='cuda fp16')
 from rwkv.utils import PIPELINE, PIPELINE_ARGS